def training_step(self, batch, batch_idx):
trainParams = {"spaceIx": args["CHAR_TO_INDEX"][" "],
"eosIx": args["CHAR_TO_INDEX"]["<EOS>"]}
inputBatch, targetBatch, inputLenBatch, targetLenBatch = batch
inputBatch, targetBatch = inputBatch.float(), targetBatch.int()
inputLenBatch, targetLenBatch = inputLenBatch.int(), targetLenBatch.int()
outputBatch = self.model(inputBatch)
with torch.backends.cudnn.flags(enabled=False):
loss = self.loss_fn(outputBatch, targetBatch, inputLenBatch, targetLenBatch)
trainingLoss = loss
predictionBatch, predictionLenBatch = ctc_greedy_decode(outputBatch.detach(),
inputLenBatch,
trainParams["eosIx"])
trainingCER = compute_cer(predictionBatch,
targetBatch,
predictionLenBatch,
targetLenBatch)
trainingWER = compute_wer(predictionBatch,
targetBatch,
predictionLenBatch,
targetLenBatch,
trainParams["spaceIx"])
self.log('train_loss', trainingLoss, prog_bar=True)
self.log('train_wer', trainingWER, prog_bar=True)
self.log('train_cer', trainingCER, prog_bar=True)
return trainingLoss
def validation_step(self, batch, batch_idx):
evalParams = {"decodeScheme": "greedy",
"spaceIx": args["CHAR_TO_INDEX"][" "],
"eosIx": args["CHAR_TO_INDEX"]["<EOS>"]}
inputBatch, targetBatch, inputLenBatch, targetLenBatch = batch
inputBatch, targetBatch = inputBatch.float(), targetBatch.int()
inputLenBatch, targetLenBatch = inputLenBatch.int(), targetLenBatch.int()
outputBatch = self.model(inputBatch)
with torch.backends.cudnn.flags(enabled=False):
loss = self.loss_fn(outputBatch, targetBatch, inputLenBatch, targetLenBatch)
evalLoss = loss
if evalParams["decodeScheme"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(outputBatch,
inputLenBatch,
evalParams["eosIx"])
elif evalParams["decodeScheme"] == "search":
predictionBatch, predictionLenBatch = ctc_search_decode(outputBatch,
inputLenBatch,
evalParams["beamSearchParams"],
evalParams["spaceIx"],
evalParams["eosIx"],
evalParams["lm"])
else:
print("Invalid Decode Scheme")
exit()
evalCER = compute_cer(predictionBatch,
targetBatch,
predictionLenBatch,
targetLenBatch)
evalWER = compute_wer(predictionBatch,
targetBatch,
predictionLenBatch,
targetLenBatch,
evalParams["spaceIx"])
self.log('val_loss', evalLoss, prog_bar=True)
self.log('val_wer', evalWER, prog_bar=True)
self.log('val_cer', evalCER, prog_bar=True)
return evalLoss
def compute_cer_checker():
preds = ["SOMETIN' ' NEDSS~", " ALRIT ~", "CHEK DON~", "EXACT SAME~"]
trgts = ["SOMETHING NEEDS~", "ALRIGHT~", "CHECK DONE~", "EXACT SAME~"]
predLens = [18, 10, 9, 11]
trgtLens = [16, 8, 11, 11]
predIxs = list()
for n in range(len(preds)):
predIx = list()
for i in range(len(preds[n])):
char = preds[n][i]
if char == "~":
ix = args["CHAR_TO_INDEX"]["<EOS>"]
else:
ix = args["CHAR_TO_INDEX"][char]
predIx.append(ix)
predIxs.extend(predIx)
trgtIxs = list()
for n in range(len(trgts)):
trgtIx = list()
for i in range(len(trgts[n])):
char = trgts[n][i]
if char == "~":
ix = args["CHAR_TO_INDEX"]["<EOS>"]
else:
ix = args["CHAR_TO_INDEX"][char]
trgtIx.append(ix)
trgtIxs.extend(trgtIx)
predictionBatch = torch.tensor(predIxs)
targetBatch = torch.tensor(trgtIxs)
predictionLenBatch = torch.tensor(predLens)
targetLenBatch = torch.tensor(trgtLens)
print(
compute_cer(predictionBatch, targetBatch, predictionLenBatch,
targetLenBatch))
return
#obtaining the prediction using CTC deocder
if args["TEST_DEMO_DECODING"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(outputBatch, inputLenBatch, args["CHAR_TO_INDEX"]["<EOS>"])
elif args["TEST_DEMO_DECODING"] == "search":
beamSearchParams = {"beamWidth":args["BEAM_WIDTH"], "alpha":args["LM_WEIGHT_ALPHA"], "beta":args["LENGTH_PENALTY_BETA"],
"threshProb":args["THRESH_PROBABILITY"]}
predictionBatch, predictionLenBatch = ctc_search_decode(outputBatch, inputLenBatch, beamSearchParams,
args["CHAR_TO_INDEX"][" "], args["CHAR_TO_INDEX"]["<EOS>"], lm)
else:
print("Invalid Decode Scheme")
exit()
#computing CER and WER
cer = compute_cer(predictionBatch, targetBatch, predictionLenBatch, targetLenBatch)
wer = compute_wer(predictionBatch, targetBatch, predictionLenBatch, targetLenBatch, args["CHAR_TO_INDEX"][" "])
#converting character indices back to characters
pred = predictionBatch[:][:-1]
trgt = targetBatch[:][:-1]
pred = "".join([args["INDEX_TO_CHAR"][ix] for ix in pred.tolist()])
trgt = "".join([args["INDEX_TO_CHAR"][ix] for ix in trgt.tolist()])
#printing the predictions
print("File: %s" %(file))
print("Prediction: %s" %(pred))
print("Target: %s" %(trgt))
print("CER: %.3f WER: %.3f" %(cer, wer))
print("\n")
